| """Snap + merge: snap baseline verts toward v9, append unmatched v9 verts. |
| |
| Best params from sweep (170 val scenes): |
| weight=0.80, radius=2.0m, greedy -> mean HSS 0.4151 (best of weight sweep 0.6-0.9) |
| """ |
| import numpy as np |
| from scipy.spatial.distance import cdist |
|
|
| SNAP_WEIGHT = 0.80 |
| SNAP_DIST = 2.0 |
|
|
|
|
| def snap_midpoint_plus_unmatched(bl_v, bl_e, v9_v, |
| weight=SNAP_WEIGHT, max_dist=SNAP_DIST): |
| """For each baseline vert, snap weight fraction toward its nearest v9 vert |
| if within max_dist. Append unmatched v9 verts as extra vertices. |
| |
| Returns (new_vertices, edges) — edges still index the original baseline vertices. |
| """ |
| bl_v = np.asarray(bl_v, dtype=np.float64).copy() |
| v9_v = np.asarray(v9_v, dtype=np.float64) |
| bl_e = np.asarray(bl_e, dtype=np.int64) |
| if len(v9_v) == 0: |
| return bl_v, bl_e |
| dists = cdist(bl_v, v9_v) |
| nearest = dists.argmin(axis=1) |
| near_dist = dists[np.arange(len(bl_v)), nearest] |
| snap_mask = near_dist < max_dist |
| bl_v[snap_mask] = ((1 - weight) * bl_v[snap_mask] |
| + weight * v9_v[nearest[snap_mask]]) |
| claimed = set(nearest[snap_mask].tolist()) |
| unmatched = [v9_v[i] for i in range(len(v9_v)) if i not in claimed] |
| if unmatched: |
| bl_v = np.concatenate([bl_v, np.array(unmatched)], axis=0) |
| return bl_v, bl_e |
|
|
